stable-worldmodel

Why we included this project

World model research usually means rewriting the same glue code each time: environment wrappers, data collection, training loops, and the planner you evaluate against. stable-worldmodel pulls those pieces behind one Python interface, so you can go from gathering demonstrations to training a model to scoring it with model-predictive control without rebuilding the pipeline. It does not pick a model for you. Instead it ships reference baselines and planning solvers, so your own architecture slots in and results stay comparable across benchmarks like PushT, Two-Room, DeepMind Control, and Atari. The data registry handles several formats, including Lance, HDF5, video, and LeRobot, which matters when data moves between local disks and cloud storage.